I think backprop can be made to work online and the cache gets instant feedback on its predictions, so I see no reason why this couldn’t be made to work. In som

I think backprop can be made to work online and the cache gets instant feedback on its predictions, so I see no reason why this couldn’t be made to work. In some ways their conditional probability estimates are one form of classifier. It would be really interesting also to see how this idea plays with specialized NN training hardware like the TPU, it would certainly make the computation easier.